My work focuses on detecting manipulated and AI-generated images and videos, with a focus on the edits that matter. Most benchmarks treat every manipulation as equally important. I care about the semantically meaningful changes that alter what an image or video says — these are the manipulations most likely to mislead people or cause harm. That work includes building manipulation pipelines for forensic benchmarks and testing whether multimodal, forensically adapted LLMs can localize and explain forgeries. More recently I’ve also studied the human side: what forensic evidence helps people judge an image, and how to turn a detection signal into something an end user can act on.
A detector that does well on a benchmark but misses the edits that change meaning, or gives output nobody can interpret, hasn’t solved the problem it was built for. I want to find where that gap opens up, and to close it with better data, better models, and better ways of presenting evidence.